为什么需要离散优化（discrete optimization）？

最新推荐文章于 2024-11-14 22:16:08 发布

原创

最新推荐文章于 2024-11-14 22:16:08 发布 · 1.6k 阅读

2 ·

CC 4.0 BY-SA版权

文章标签：

#discrete-reg

本文探讨了在肺部CT中，通过马尔可夫随机场的离散优化策略避免大变形配准中的局部最小问题。作者提出一种密集采样方法，结合B-spline变换模型，有效处理呼吸引起的显著变形，并利用MRF标记控制位移。文章还提及了最小生成树在配准流程中的应用，以及降低计算复杂性的策略。

一、MRF-based deformable registration and ventilation estimation of lung CT

1. Introduction:

A. Large motion of small features

Motion within the lungs can often be larger than the scale of the features (vessels and airways). This may cause a registration algorithm to get trapped in a local minimum, and may lead to an erroneous registration. Local minima are frequently encountered in lung registration.
…
An approach to avoid local minima is the use of discrete optimisation, which is usually formulated on a Markov random field (MRF). Discrete optimisation offers numerous advantages, in particular a greater control over the displacement space, to overcome these limitations.

深度学习方法在大变形配准下表现很差，比如肺部的呼吸运动导致的大变形往往比深度学习提取的特征（比如血管和气道）要大，这时候深度学习就容易陷入局部最优的陷阱。

一个避免局部最优陷阱的方法就是使